Compilers

TensorFlow has accumulated several compilers and lowering paths. They all turn TF graphs into something more efficient — XLA HLO, MLIR dialects, TFLite flatbuffers, or static C++ binaries.

Page	What it covers	Code under
XLA / tf2xla / JIT / AOT	TF→HLO lowering, JIT auto-clustering, ahead-of-time tfcompile	tensorflow/compiler/tf2xla/, jit/, aot/
MLIR	The MLIR dialects TF defines and the passes between them	tensorflow/compiler/mlir/, tensorflow/core/ir/, tensorflow/core/transforms/
TensorRT integration	NVIDIA TensorRT subgraph fusion	tensorflow/compiler/tf2tensorrt/

How they connect

graph LR
    Graph[TF Graph or FuncGraph]
    Grappler[Grappler]
    JIT[JIT auto-clustering]
    TF2XLA[tf2xla]
    MLIR[MLIR pipeline tf -> mhlo -> stablehlo]
    HLO[XLA HLO]
    XLA[XLA compiler in openxla/xla]
    Device[CPU/GPU/TPU exec]
    AOT[tfcompile AOT]
    Static[Static C++ binary]
    TF2TRT[tf2tensorrt]
    TRT[TensorRT engine]
    TFLite[TFLite flatbuffer]

    Graph --> Grappler
    Grappler --> JIT
    JIT -- chosen subgraphs --> TF2XLA
    Graph --> MLIR
    MLIR --> HLO
    TF2XLA --> HLO
    HLO --> XLA
    XLA --> Device
    Graph --> AOT
    AOT --> XLA
    AOT --> Static
    Graph --> TF2TRT
    TF2TRT --> TRT
    MLIR -- tfl dialect --> TFLite

What lives where

• The XLA compiler core (HLO IR, optimization passes, codegen) was extracted to openxla/xla in 2023. This repo only contains the bridge code (TF↔XLA) under tensorflow/compiler/.
• MLIR is used by both the new TFLite converter and the new TF compiler stack. Dialects and passes live under tensorflow/compiler/mlir/.
• TensorRT (NVIDIA's inference compiler) integration is under tensorflow/compiler/tf2tensorrt/ and produces a TRTEngineOp that runs the engine inline in a TF graph.
• AOT (tfcompile) generates static C++ libraries for an XLA-compiled function — the path used to ship inference models without bringing the whole TF runtime.

Compiler vs runtime

The boundary is fuzzy:

• Compilers transform graphs offline (or at first call) and emit some artifact (HLO, TF executable, TFLite flatbuffer, static lib).
• Runtimes then execute the artifact (core-runtime, TFRT, TFLite, or XLA's runtime in openxla/xla).

A tf.function(jit_compile=True) call straddles both: the compiler runs at first call, the runtime runs the resulting executable on subsequent calls.